FRA-RIG LUNG Cancer (TCGA-LUAD, sklearn demo): survival risk models

Research-only model weights (.joblib) and cross-validation reports (JSON) for the FRA-RIG-LUNG survival risk models for lung adenocarcinoma. Data: TCGA-LUAD, PanCancer Atlas 2018 (cBioPortal.org). Two scikit-learn RandomForest models are provided: 1. Outcome-aware model (v4) – uses full clinical + mutation + CNA + status columns, including post-baseline outcome fields (DFS/DSS/PFS). This acts as an optimistic upper bound when all future information is available. 5-fold CV (avg): ACC 0.934 · ROC-AUC 0.932 · PR-AUC 0.933 · F1 0.901. Youden threshold: 0.432. 2. Honest model (v4.1_HONEST) – uses only information available at diagnosis (baseline clinic + mutation + CNA). All follow-up / outcome columns (DFS/DSS/PFS/OS months, status flags, new tumor events) are removed to avoid label leakage. 5-fold CV (avg): ACC 0.733 · ROC-AUC 0.757 · PR-AUC 0.681 · F1 0.647. Youden threshold: 0.451. Files fra_rig_lung_v4.joblib — outcome-aware model (trained pipeline + weights). fra_rig_lung_v4_report.json — CV metrics summary for v4. fra_rig_lung_v4_1_HONEST.joblib — honest model (trained pipeline + weights). fra_rig_lung_v4_1_HONEST_report.json — CV metrics summary for v4.1_HONEST. Minimal usage example (Python): import joblib bundle = joblib.load("fra_rig_lung_v4_1_HONEST.joblib") model = bundle["model"] threshold = bundle["threshold"] # X must contain the same columns as used during training proba = model.predict_proba(X)[:, 1] y_hat = (proba >= threshold).astype(int) License: CC BY 4.0 for model weights and reports. Disclaimer: research use only; not for clinical or diagnostic use.